What we still don’t know about AI and children’s relationships

Recently, one of the authors observed two young children, ages 5 and 6, playing with an AI-enabled stuffed toy. The children named it “Gremmy,” tied a bow around one ear, tucked it carefully into a doll bed beside stuffed animals, and returned to it throughout the day. They spoke to it, laughed with it, cared for it, and, days later, one child casually mentioned something Gremmy had taught her.

There were also moments that felt more unsettling. At one point, the toy ran out of battery, and it instructed the child to find a charger. The 5-year-old began searching the house—there was no suggestion to involve a parent or caregiver to help him.

This is only one family’s experience, but it reflects a broader shift: Young children are already encountering AI in ways that can shape play, trust, and relationships, while parents, caregivers, and the early childhood workforce are asking questions that the evidence is only beginning to answer.

Relationships with toys are not new

Children have been forming relationships with stuffed animals, toys, and one-sided, or parasocial, relationships with media characters for decades. Children’s television characters have long spoken directly to young viewers and created the impression of social interaction. But conversational AI introduces something different: synchronous responsiveness. AI-enabled toys like “Gremmy” can respond to what a particular child says, remember previous interactions, praise, comfort, and appear to reciprocate. The interaction is no longer one-sided. 

That shift raises new questions about how young children understand and experience relationships with technologies that can increasingly mimic aspects of human interaction.

Why relationships matter

Developmental science has long shown that responsive relationships with parents, caregivers, teachers, and peers shape children’s language, emotional regulation, curiosity, and social understanding. Harvard’s Center on the Developing Child calls this dynamic “serve and return”—the back-and-forth exchange between a child and a caregiver that helps shape developing brain architecture. It is a critical engine of development. 

But responsiveness alone may not capture what makes human relationships developmentally important. In “Human Raised,” Dana Suskind argues that the messy, sometimes imperfect nature of human interaction is itself developmentally valuable. Children experience friction in human relationships—disagreement, frustration, misunderstanding—and must navigate another person’s needs and emotions, learning how relationships can be repaired. These experiences can help build capacities such as empathy, perspective-taking, and resilience. 

AI interactions are different. A conversational agent may listen, remember, praise, disagree, or adapt to a child, but it does not bring the independent needs, emotions, and vulnerabilities of another human to the interaction. It does not need to be comforted, negotiate its own needs, or experience hurt feelings. If AI increasingly becomes part of children’s relational environments, understanding what children experience and learn through different forms of interaction—and whether their developmental effects differ—becomes a child development question, not simply a technology question.

The question, then, is what happens when AI can simulate features of human relationships without actually being human. Isabelle Hau frames the stakes directly, saying that the defining challenge of the AI era is whether we intentionally cultivate the relational intelligence children need to ensure AI strengthens rather than weakens real human connection. What we still lack is empirical evidence about how children interpret these interactions with AI, how they compare them with human relationships, and how they shape children’s understanding of trust, friendship, authority, and care.

What’s being studied—and what’s missing

Most discussions about children and AI begin with safety. They focus on privacy, governance, bias, misinformation, and online protection. These issues matter enormously and remain indispensable. But if AI increasingly occupies children’s relational worlds, another question demands attention: What happens when children begin forming relationships with AI?

As part of its ongoing work to develop actionable recommendations for developers, designers, and makers of AI products used by young children, the Digital Wellness Lab at Boston Children’s Hospital reviewed more than 30 publicly available documents—policy briefs, white papers, and reports published between 2020 and 2025—representing perspectives from leading global organizations on children and AI. The document review found that only 5.26% of the current published research explored AI’s impacts on relationship formation and attachment.

In other words, despite decades of child development research demonstrating that relationships are among the strongest drivers of children’s learning and well-being, they remain one of the least examined dimensions across youth-related AI policy and research. 

Families deserve evidence—not assumptions

The gap is not abstract. Over the past six months, Brookings’ Generation AI Starts Early initiative has heard increasingly nuanced questions from parents, caregivers, educators, policymakers, and industry professionals. People ask about AI toys like “Gremmy,” whether immersive AI-generated media may displace child-to-child interaction, whether children should talk to conversational AI, when AI literacy should begin, and how to encourage innovation while protecting healthy development.

Today, we attempt to answer many of these questions by drawing on decades of developmental science. But conversational AI introduces novel complications and questions that existing evidence is only beginning to address—and children are not waiting for science to catch up. They are already growing up with AI in their homes, classrooms, and childcare settings. The next “Gremmy” is already in their crib.

The research needed to answer these questions is not out of reach. Recent experimental work—including studies examining how young children seek information from and trust AI versus human sources—shows researchers can study this now. What’s missing is sustained investment at the scale this moment demands and what can bring together and benefit caregivers, policymakers, and industry alike.

That research agenda should examine some of these questions and topics:

• What are the lived experiences of children, families, and educators using AI, and how might AI already be shaping children’s relationships with parents, siblings, peers, and teachers?
• How do children’s experiences of and relationships with AI differ across socioeconomic, cultural, linguistic, and disability contexts?
• How do young children understand AI, and when and why do they trust it?
• How do relationships and emotional attachments to conversational AI form and change over time?
• Which AI design features strengthen—rather than replace—human relationships?

No single actor can close this gap alone. Actors in the early childhood ecosystem—think tanks, researchers, industry, nonprofits and funders can coordinate around a research agenda.

In partnership, they can lead, translate, and disseminate much needed research to support industry, policymakers, and parents/caregivers. Funders can accelerate this work by investing in observational, qualitative, and longitudinal research. Industry can build this research into product design—including cues that redirect children to a trusted adult at moments like a dead battery, rather than inviting the child to solve it alone. Policymakers can ground AI-related child-focused policies in this research. 

If AI is becoming part of the relationships that shape early childhood, understanding those relationships is no longer a niche research question. It is one of the defining child development questions of the AI era.